VAE Loader
Grey, washed-out images? This is the node that fixes them
- vae
Every Stable Diffusion user eventually hits the grey images. Not washed out, not low contrast - literally grey, like the color was drained out of the picture. That's the VAE failing, and TMMVaeLoader is the node you'd reach for in a graph built by the Tiny Model Manager dashboard.
What it does
vae_name(dropdown) - every file inComfyUI/models/vae.vae(VAE) - wire intoVAE Decode(or into the VAE input a loader exposes) so the sampler's latent actually turns into a colored image.
The VAE is the component that converts between latent space (the compressed representation the diffusion model denoises in) and pixel space (what you see). It's the last mile of every generation, and when it's missing or wrong, the decode comes out grey and flat. Load the right VAE and the color comes back.
Baked vs. separate - the context that decides whether you need this at all
Here's the thing: many checkpoints bake their VAE right into the checkpoint file, so a standalone VAE loader is unnecessary - the vae output on your checkpoint loader is already doing this job. Others (a good chunk of the SD 1.5 era) expect you to load one separately, and those are the ones that produce the grey disaster if you don't. So this node isn't a mandatory part of every graph; it's the tool for when your model card says "needs a VAE" or the output looks like a rainy day.
The gotcha that matters: don't mix architectures
A VAE is not a universal component. SD 1.5, SDXL, and Flux use different latent channel counts, and a VAE from one architecture on another produces noise or flat color - sometimes "wrong" in a way that looks almost right, which is worse than obvious garbage. The dashboard's model cards store base-model metadata for the very purpose of not doing this. If you're manually hunting VAEs, remember the classic 2023 fix was vae-ft-mse-840000-ema-pruned for SD 1.5; SDXL and Flux bring their own and generally don't want yours.
How it works
comfy.sd.VAE(ckpt_path=...) - the exact same call as ComfyUI's core VAELoader. One file in, one VAE object out, no guess-config needed. It's a thin wrapper whose real reason to exist is being the matching node for the dashboard's "+" button on a VAE card.
Install
Same one-time setup for the whole suite:
cd ComfyUI/custom_nodes
git clone https://github.com/Zellione/comfyui-tiny-model-manager
cd comfyui-tiny-model-manager
pip install -r requirements.txt
or search Tiny Model Manager in ComfyUI Manager, restart, and the nodes live under tiny-model-manager in the node menu. Light Python deps, prebuilt dashboard, nothing heavy.
If your checkpoint bakes a VAE, you'll never touch this node and that's fine. The day you load a model that doesn't, you'll be glad the dashboard put a VAE Loader two clicks away.
Inputs (1)
| Name | Type | Default | Description |
|---|---|---|---|
| vae_name | COMBO | 0 options: |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| vae | VAE | — |